On 18 October 2016 at 13:09, Peter Zijlstra <[email protected]> wrote: > On Wed, Oct 12, 2016 at 09:41:36AM +0200, Vincent Guittot wrote: > >> ok. In fact, I have noticed another regression with tip/sched/core and >> hackbench while looking at yours. >> I have bisect to : >> 10e2f1acd0 ("sched/core: Rewrite and improve select_idle_siblings") >> >> hackbench -P -g 1 >> >> v4.8 tip/sched/core tip/sched/core+revert 10e2f1acd010 >> and 1b568f0aabf2 >> min 0.051 0,052 0.049 >> avg 0.057(0%) 0,062(-7%) 0.056(+1%) >> max 0.070 0,073 0.067 >> stdev +/-8% +/-10% +/-9% >> >> The issue seems to be that it prevents some migration at wake up at >> the end of hackbench test so we have last tasks that compete for the >> same CPU whereas other CPUs are idle in the same MC domain. I haven't >> to look more deeply which part of the patch do the regression yet > > So select_idle_cpu(), which does the LLC wide CPU scan is now throttled > by a comparison between avg_cost and avg_idle; where avg_cost is a > historical measure of how costly it was to scan the entire LLC domain > and avg_idle is our current idle time guestimate (also a historical > average). > > The problem was that a number of workloads were spending quite a lot of > time here scanning CPUs while they could be doing useful work (esp. > since newer parts have silly amounts of CPUs per LLC).
make sense > > The toggle is a heuristic with a random number in.. we could see if > there's anything better we can do. I know some people take the toggle > out entirely, but that will regress other workloads. ok. so removing the toggle fixes the problem in this test case too. may be we can take into account the sd_llc_size in the toggle
